AlphaNova Hits 10,000 Users
AlphaNova Hits 10,000 Users: Here's What We've Learnt from Our Users
We crossed 10,000 registered users last week.
Ten thousand people, and still no tidy way to describe the group. Quants. Kagglers. Mathematicians. ML engineers. Algo traders. PhD researchers. Students. Career switchers. People who just like a hard problem and don't care what it's called.
Thank you to every one of them.
Because the more time we spend inside this community, the more obvious one thing becomes: you cannot tell from the outside who is going to find signal.
What the Leaderboard Keeps Telling Us
We've now run multiple cycles and published profiles of some of the people who made it through.
A International Quant Championship 24 National Finalist succeeds in one cycle. Someone with no formal coding background enters the top 3 in the next.
A psychology graduate who started programming at nine builds a research engineering workflow that rivals what we see from professional desks. A fraud-detection engineer builds a system to catch his own overfitting—and then catches himself overfitting the meta-analysis.
A freshman places top‑1% in a global quant championship. A high school graduate with a $20/month Claude subscription cracks the top ten.
Someone from Singapore balancing a double specialisation, two minors, and CFA Level I reaches for Ridge regression rather than a neural network. Someone from India deliberately sacrifices Sharpe to reduce correlation with the crowd—and jumps from rank 167 to rank 6.
As our co‑founder Marc Nunes puts it:
"Because the next great source of alpha may not emerge from the person with the most impressive résumé, the most elegant theory, or even the most convincing explanation. It may emerge from a machine, a process, or an unlikely combination of ideas that simply discovers something true before everyone else does."
Ten Thousand People From Every Part of the World
Part of what makes this community interesting is where it comes from. Our users are spread across:
🇮🇳 India · 🇯🇵 Japan · 🇺🇸 United States · 🇬🇧 United Kingdom · 🇫🇷 France · 🇩🇪 Germany · 🇧🇩 Bangladesh · 🇵🇰 Pakistan · 🇨🇦 Canada · 🇸🇬 Singapore · 🇭🇰 Hong Kong · 🇨🇳 China — and plenty more countries beyond that.
This matters more than it might sound. Crowdsourced signal discovery only works when the crowd is genuinely diverse—when people approach the same problem from different angles, with different priors, and different intuitions about what "looks right." A monoculture of identical approaches produces identical signals, which is exactly what our greedy quality selection is designed to filter out.
Geographic diversity isn't a nice-to-have. It's a structural advantage. As our research into signal libraries and correlation geometry shows, the space of genuinely separated signals is far larger than it first appears—and the variety of histories people can bring is what makes the collection stronger. That research is built on three papers recently published by Marc: Signal Correlation, IC, and PnL Dependence, Large Signal Libraries: Equal-Weight Limits and the Divergent Spectra of Signals and PnL, and Separated Signal Libraries: Packing, Saturation, and Joint Spectral Limits.
Three Things We've Learned from 10,000 Users
1. Titles tell you nothing
Across the profiles we've published, the range of backgrounds is wide: mechanical engineering, psychology, economics, electrical engineering, high school, PhD programmes, career switchers. What they share isn't a credential. It's the willingness to test something properly and the tolerance for being wrong.
2. The infrastructure around the model matters more than the model
This has come up repeatedly. Mathurin built a fast local evaluator and a surrogate model of the validation-to-test haircut. Alok did the same thing independently and calibrated his local Sharpe to the server at roughly 0.52x. Ismam retired an entire architecture rather than tune it into a prettier backtest.
None of them succeeded because of a clever model. They succeeded because of the systems they built around it.
3. AI narrows the implementation gap—and widens the judgment gap
Taro proved you don't need to write code to compete. Braxton uses AI agents connected to his own research notes. Saksham generated 28 candidate signals in two weeks with AI assistance but re‑ran every result himself before trusting it.
The pattern is consistent: AI makes experimentation cheap. That means the scarce resource is now judgment—knowing what to test, what to believe, and what to throw away.
Our Biweeklies Are Open
If you've been reading these profiles and wondering whether you'd fit in: the answer is almost certainly yes, and the only way to find out is to submit something.
Our biweekly competitions are live. Pure Python, obfuscated data, walk‑forward evaluation, no staking. Cash prizes paid in stablecoins or to your bank account. Your IP stays yours.
To the 10,000 who are already here: thank you. To the next 10,000: the leaderboard doesn't care where you're from or what you studied. It only cares whether your signal survives.
See you on the board.
Further reading from the series: